Acident report information extraction model based on ALBERT-BiGRU-CRF

As global traffic accidents occur frequently, comprehensively understanding relevant information is particularly important. This not only helps identify potential safety hazards but also improves traffic infrastructure and management, thereby reducing the incidence of accidents. However, the effective extraction of key data such as time, location, and casualties from existing traffic accident texts still faces challenges and to address the low accuracy of traditional text extraction methods in extracting traffic accident information, this paper employs deep learning models for information extraction, aiming to mine effective information from traffic accident texts.First, we customized methods and identification standards for named entity recognition in traffic accidents to ensure the accuracy and consistency of the extraction. To mitigate the impact of data sparsity on model performance, we introduced a lightweight pre-trained model, ALBERT. This model generates high-quality embeddings that serve as input feature vectors for the Bi-directional Gated Recurrent Unit Conditional Random Field (BiGRU-CRF) model, enhancing the efficiency of entity feature extraction. Finally, we compared the extraction performance of this model with four other mainstream deep learning models. The results indicate that this model achieves an average accuracy of 0.94 in extracting information from traffic accident reports, demonstrating its strong extraction performance.